How do you write content AI can cite?

Writing rules for citable content: the question in the heading, the answer in the first paragraph and figures with source and date.

← back to home

AI visibility // writing
citable.md

$ write --rule=answer-first

How do you write content AI can cite?

Content AI can cite answers the question in the heading directly in the first paragraph, in passages that read on their own, separate from the rest. Work with clear headings that match real search questions, tables for comparisons, an FAQ, figures with source and date, and a named author. These are writing rules, not tricks; they also make your text better for readers.

DD DataDrift Digital • 16 September 2026 • 5 min

An AI system compiling an answer isn't looking for an attractive page but a usable passage: a piece of text that answers the question without needing the rest of the page. Writing for citability is therefore mainly about creating such passages. The rules below do exactly that.

01 / the answer firstThe question in the heading, the answer below

The basic rule: put your reader's question literally in the heading, and give the full answer in the first paragraph below it, in forty to sixty words. Not the lead-up, not the context, the answer. The explanation, the nuance and the examples come afterwards.

That order feels backwards to many writers, because it gives away the tension a text would normally build up. But a searcher isn't a reader of a story: they want to know first whether they're in the right place. And an AI system selecting passages picks the block where question and answer sit together, not the page where the answer is hidden in paragraph eight.

The test is simple: read only your heading and your first paragraph, and judge whether someone is helped by it. If not, your piece starts too late. This article does the same itself; the answer to the title question was at the top, and everything after that is depth.

02 / the building blocksGive each type of content its own form

After the first paragraph, form determines how readable the rest is. The rule of thumb: each type of content has a fixed form that can be quoted.

Type of contentForm that can be cited
Concept or definitionOne paragraph that stands on its own
Way of working or processNumbered steps
ComparisonTable with the differences named
Frequently asked questionsFAQ with independently readable answers

The key test per block: is this piece understandable to someone who only sees this piece. If so, it's citable. If not, it leans on context an AI system won't supply. So avoid references like "as described above" within a block; each block carries its own context, or it won't be used.

03 / the substantiationFigures, sources and a name

Between two passages that say the same thing, the one that substantiates it wins. State figures with their source and year, refer to the original research rather than a summary of it, and put a named author under the piece. Also show when the text was last updated: recency counts, and a dated piece loses out to a maintained one.

Substantiation also means being honest about what you don't know. An estimate labelled as an estimate is citable; an estimate written up as fact gets found out sooner or later and drags down the rest of the page. So write "we see in our own projects" rather than a percentage without provenance.

Structure is for the machine, text is for the human.

04 / the pitfallWriting for the machine backfires

The temptation is to overdo these rules: every sentence a keyword, every paragraph an optimised little block. That demonstrably backfires. The GEO research from Princeton (KDD 2024) measured that keyword stuffing made performance 10 percent worse on Perplexity. Google also explicitly warns against chopping content into small pieces for AI and against writing separate texts for AI systems.

So the measure remains the reader. A text a human won't finish reading isn't improved by a machine. The structure in this piece organises what you have to say; it doesn't replace having something to say. How these writing rules fit into the broader approach is set out in how you get mentioned in ChatGPT answers.

05 / the limitWhen writing rules won't save you

These rules only help if the rest is in place. If your site is technically unreadable for crawlers, even the best-structured text will never be found; check that first. Additional technology such as structured data can then widen the difference, which is covered in does schema markup help AI visibility.

And if nobody is searching for what you write, form changes nothing about that. Write about the questions your customers actually ask you, in their words. If you don't have those questions clear, that's your first job, not the writing.

The rules in one list, as a checklist for every piece you publish:

  • The question literally in the heading, the full answer in the first paragraph of forty to sixty words.
  • Each block independently readable, without references like "as described above".
  • Comparisons in a table, processes in numbered steps, concepts in one self-contained paragraph.
  • Figures with source and year; estimates labelled as estimates.
  • A named author and a visible update date.
  • No keyword stuffing and no separate texts for AI systems.
  • Checked in advance that crawlers can read the page.
Frequently asked questions
What makes a text citable for AI?+
Self-contained passages: the question in the heading, the full answer in the first paragraph of forty to sixty words, and blocks that are understandable without the rest of the page. On top of that, substantiation with sources, a named author and a visible recency date all count.
How long should the answer in the first paragraph be?+
Forty to sixty words is the workable measure: long enough to answer the question fully, short enough to be quoted in its entirety. The paragraph should be independently readable, so without references to something explained further on.
When do you use a table instead of text?+
For every comparison. Two options side by side, differences per feature, pros and cons: a human reads that faster in a table, and an AI system can cite from it cell by cell. For processes a numbered list works better, and for concepts a self-contained paragraph.
Is writing for AI different from writing for humans?+
No, and that's the core of it. Google explicitly advises against separate texts for AI systems, and stuffed texts reduce visibility: the Princeton research (KDD 2024) measured that keyword stuffing made performance 10 percent worse on Perplexity. Structure organises the text for the machine; the content remains for the reader.
Does an FAQ block really help AI visibility?+
Yes, provided the answers are independently readable. An FAQ puts question and answer literally together, which is exactly the format in which AI systems select passages. Phrase the questions the way your customers ask them, not the way you refer to them internally, and keep each answer complete in itself.
baseline measurement · fixed fee · no obligations

Is your current content cited anywhere in an AI answer?

The scan

This is the baseline measurement: fixed search queries on ChatGPT, Gemini, Claude, Perplexity and Google. Current prices are on the pricing page. Even without a follow-up, you keep the report.

→ The scan · pricing

This text was produced with AI assistance and checked and approved by a human before publication.